code/scripts
| script | purpose | device |
|---|---|---|
container_smoke.sh |
serve ONE serve profile of the built package (--profile NAME; this package has only the default), run server/smoke_test.py against it (asserts ETH dispatch, the 12x10 grid and the profile's pins; compares the shipped sample's output with its stored CPU reference under the PLAN box agreement), keep the evidence (container log, /info, the /predict output and the result, in logs/smoke/ of the repo or --log-dir DIR), always stop it; one bin/devrun -t 3600 -k 150 window |
yes |
bench.py |
stage bench of warm forwards (ttaw.profiling.StageBench): host preprocess / host->ttnn / H2D / trace replay / D2H / host BEVPool / postprocess / e2e p50-p99 per input (--inputs synthetic[=preset],pandaset[=dir],nuscenes[=manifest],sample=<manifest>), device-only replays and back-to-back frames, AICLK, a PandaSet agreement check (workspace data); --dispatch / --num-cqs for the D14 matrix (OPT_BASELINE.md) |
yes, via bin/devrun |
profile_ops.py |
device profile: each trace variant replayed between its own signposts (encode / decode / push / fill, --rounds), --eager adds one eager encode + decode with module-stage signposts; --graph-report DIR writes a ttnn-visualizer memory / graph report instead (python -m tracy -r -p -v --no-web-server ..., OPT_BASELINE.md) |
yes |
profile_summary.py |
summary of a profile_ops.py ops CSV: programs, kernel sum, op-to-op gaps and span per segment, per-module stages (eager labels aligned onto the traced ops), top ops and gaps |
no |
probe_mm.py, probe_conv.py, probe_l1res.py, debug_l1res.py, probe_upk.py, probe_conv512.py |
optimization probes (OPT_REPORT rounds 2-6): 1x1 linear configs, per-conv Conv2dConfig candidates, the L1-resident bottleneck (tt.l1res) vs the round-3 path, and its per-block bit check on the real weights; the upk up-sampling kernel (bring-up shapes, host model / float64 / grid_sample errors, trace replay); transposed (COL_MAJOR) block sharding of one conv on chosen grids (DRAM in / out and chained); bit identity + device time per candidate |
yes |
run_public_tt.py |
public-data manifests (research/bevdet/public_data of the tt-models workspace: nuScenes v1.0-mini, PandaSet) through the API, results in the CPU reference's format for the agreement, the mAP / NDS scripts and make_demo.py (--out-dir) |
yes |
make_sample.py |
generates the shipped sample tt_bevdet/samples/synthetic_6cam* (a synthetic six-camera test pattern), its preset calib/synthetic_6cam.json, the stored CPU reference and the sparse reference heads of the host tests (optimise, then finalise; research venv, CPU) |
no |
make_demo.py |
renders media/ (camera mosaics with the TT boxes, BEV plots, TT vs CPU) from run_public_tt.py outputs and the TT output of the shipped sample; writes media/ATTRIBUTION.md (tt-models workspace: datasets + the research renderer) |
no |
make_goldens.py |
per-module goldens of the TT port from the fp32 CPU reference (tt_bevdet.reference), cross-checked against the research ONNX Runtime goldens; big ones to research/bevdet/goldens (never shipped); --sets pandaset --no-big writes the PandaSet small goldens and stored reference into the staging dir (research venv, CPU) |
no |
stage_pandaset_sample.py |
(re)creates / checks the git-ignored staging_samples_pandaset/ (the PandaSet 019 f38-40 sample, prepared but not shipped until the user approves PandaSet samples) |
no |
fetch_samples.sh |
downloads the non-redistributable validation key-frame of bevdet_vendor (nuScenes sample0, sha256-checked) and points at nuScenes v1.0-mini / PandaSet |
no |
Everything here ships in code/ on the Hub and inside the image (source.extra_code lists scripts); the scripts
marked "workspace" need data of the tt-models workspace that is not redistributed.